The flat-rate AI era is over.

The flat-rate AI era is over.

For two years, AI felt like it had a fixed cost. You bought a subscription, you got a model. And as the labs dramatically increased their capabilities, the bill stayed flat, though you probably crept up your spend on everything else that runs on AI too. There was always a sense that our usage was being heavily subsidized. It was.

That subsidy era is ending, and the model providers are starting to post real profits. Anthropic is forecasting $10.9B in Q2 revenue and a $559M operating profit — years ahead of its own schedule.

As Simon Willison put it last month, AI has finally found product-market fit, not on chatbots, but on coding agents, and the infrastructure underneath it can now charge for the value it creates. Willison's own receipt makes it concrete: he burns roughly $2,180/month in API tokens while paying about $200 in consumer plans.

And yet AI is still cheap compared to the alternatives. At Valet, we see Sonnet 4.6 running in our environment for about $12/hr, far less than minimum wage in California. So there's enormous economic pressure to consume more tokens with the people we already have.

A quick reminder of what we're counting

A token is the unit of work in a large language model, roughly a word, or part of one. For most of 2024, individual users consuming tokens at scale were an edge case. Flat-rate plans worked precisely because the average user never pushed hard enough to stress the economics.

For a while, tokens-per-employee became a decent proxy for AI adoption, and companies built leaderboards to gamify it. But every game without rules has unintended consequences. Amazon found this out the hard way last week: it killed an internal leaderboard called KiroRank after employees started “tokenmaxxing”, running low-value tasks through AI agents purely to climb the rankings. An SVP had to send a memo asking people to stop using AI just for the sake of using AI. Meta had its own version, a “Claudeonomics” dashboard that ranked 85,000 workers by token consumption and clocked 60 trillion tokens in thirty days. Turns out, when you reward volume, you get volume, not value.

Token exhaustion is real

My cofounder and I regularly hit the ceiling of Claude's Max plans, especially when we're using agents well. Because an agent doesn't read one document and stop. It reads hundreds. It writes and rewrites. It calls other agents. It runs in loops. A single agentic workflow can burn through in an afternoon what a normal user spends in a month. There's a good version of this and a bad version.

The good version is intentional: You know which workflows are token-hungry, you designed them that way on purpose, and the output justifies the spend. Dan Shipper at Every just did a podcast with Lenny on his experience. He added capabilities he never had the headcount for. Agents that work around the clock. Agents that don't get tired of running six parallel analyses before the morning standup.

The bad version is accidental: You give every team the best model in every context, no guardrails, no visibility into what's being spent on what, and you incentivize raw usage. Then you get the Uber situation: they wired Claude Code into engineering, ranked teams on a usage leaderboard, and burned through their entire 2026 AI budget in four months. Per-engineer spend ran $500 to $2,000 a month. And when their COO went looking for the payoff, he couldn't find it: it was “very hard to draw a line,” he said, between the Claude Code spend and useful consumer features.

The problem was never that Uber used too much AI. It's that nobody built the architecture to know whether that usage was producing proportional value. Software engineering and ML were mature fields long before LLMs, and it turns out deterministic software is still cheap to run. The expensive part is judgment, and judgment is exactly what you shouldn't be spending frontier tokens on indiscriminately.

We're at a cost inflection point. Locking into the most expensive model for every workflow is not the right call. A tiered approach, frontier models for the highest-judgment tasks, efficient models for high-volume routine work, local models for asynchronous (slow) tasks, is where I see more teams heading.